In this paper we propose a semantic approach to automatically identify and remove stopwords from Twitter data. Unlike most existing approaches, which rely on outdated and context-insensitive stopword lists, our proposed approach considers the contextual semantics and sentiment of words in order to measure their discrimination power. Evaluation results on 6 Twitter datasets show that, removing our semantically identified stopwords from tweets, increases the binary sentiment classification performance over the classic pre-complied stopword list by 0.42% and 0.94% in accuracy and F-measure respectively. Also, our approach reduces the sentiment classifier's feature space by 48.34% and the dataset sparsity by 1.17%, on average, compared to the classic method.
The excessive presence of abbreviations and irregular words in tweets make them very
noisy, sparse and hard to extract sentiment from [
In this paper, we propose an unsupervised approach for automatically generating
context-aware stoplists for the sentiment analysis task on Twitter. Our approach captures
the contextual semantics and sentiment of words in tweets in order to calculate their
informative value. Words with low informative value are then selected as stopwords.
Contextual semantics (aka statistical semantics) are based on the proposition that meaning
can be extracted from words co-occurrences [
We evaluate our approach against the Van stoplist (so-called clasic method) using six Twitter datasets. In particular, we study how removing stopwords generated by our approach affects: (i) the level of data sparsity of the used datasets and (ii) the performance of the Maximum Entropy (MaxEnt) classifier in terms of: (a) the size of the classifier’s feature space and, (b) the classifier’s performance. Our preliminary results show that our approach outperforms the classic stopword removal method in both accuracy and F1-measure by 0.42% and 0.94% respectively. Moreover, removing our semanticallyidentified stopwords reduces the feature space by 48.34% and the dataset sparsity by 1.17%, compared to the classic method, on average.
The main principle behind our approach is that the informativeness of words in sentiment analysis relies on their semantics and sentiment within the contexts they occur. Stopwords correspond to those words of weak contextual semantics and sentiment.Therefore, our approach functions by first capturing the contextual semantics and sentiment of words and then calculating their informative values accordingly. 2.1
To capture the contextual semantics and sentiment of words, we use our previously
proposed semantic representation model SentiCircles [
In summary, the SentiCircle model extracts the contex- Y +1 tual semantics of a word from its co-occurrences with other Very Positive Positive words in a given tweet corpus. These co-occurrences are then Ci yi tironepgcroessimmepnptueltede tatrhsigeaocgnoeonomtmeexettrturiciaclicdsieerncnltteiimtiweehsnitochnofiistt.hseuInbwspeoaqrrdutiecbnuytllyaarpu,psfleoydr- -1 xi ri θmi each unique term m in a tweet collection, we build a twoidnimtheenscieonntarlegoefomtheetrciicrcclier,claen,dwehaecrhe tphoei ntetramromunids istitrueaptreed- FiVge.r1y:NeSgeantitvieCircle-1 of a Nteegramtivem. sents a context term ci (i.e., a term that occurs with m in the Stopwords region is shaded in same context). The position of ci, as illustrated in Figure 1, gray. is defined jointly by its Cartesian coordinates xi; yi as: +1 X xi = ri cos( i ) yi = ri sin( i ) Where i is the polar angle of the context term ci and its value equals to the prior sentiment of ci in a sentiment lexicon before adaptation, ri is the radius of ci and its value represents the degree of correlation (tdoc) between ci and m, and can be computed as: ri = tdoc(m; ci) = f (ci; m) log(N=Nci ) where f (ci; m) is the number of times ci occurs with m in tweets, N is the total number of terms, and Nci is the total number of terms that occur with ci. Note that all terms’ radii in the SentiCircle are normalised. Also, all angles’ values are in radian.
The trigonometric properties of the SentiCircle allow us to encode the contextual semantics of a term as sentiment orientation and sentiment strength. Y-axis defines the sentiment of the term, i.e., a positive y value denotes a positive sentiment and vice versa. The X-axis defines the sentiment strength of the term. The smaller the x value, the stronger the sentiment.1 This, in turn, divides the circle into four sentiment quadrants. Terms in the two upper quadrants have a positive sentiment (sin > 0), with upper left quadrant representing stronger positive sentiment since it has larger angle values than those in the top right quadrant. Similarly, terms in the two lower quadrants have negative sentiment values (sin < 0). Moreover, a small region called the “Neutral Region” can be defined. This region is located very close to X-axis in the “Positive” and the “Negative” quadrants only, where terms lie in this region have very weak sentiment (i.e, j j t 0). 1 This is because cos < 0 for large angles.
overall sentiment of m is by calculating the geometric median of all the points in its SentiCircle. Formally, for a given set of n points (p1; p2; :::; pn) in a SentiCirlce , the 2D geometric median g is defined as: g = arg ming2R2 Pin=1 kjpi gjj2. We call the geometric median g the SentiMedian as its position in the SentiCircle determines the total contextual-sentiment orientation and strength of m. 2.2
Stopwords in sentiment analysis are those who have weak semantics and sentiment within the context they occur. Hence, stopwords in our approach are those whose SentiMedians are located in the SentiCircle within a very small region close to the origin, as shown in Figure 1. This is because points in this region have: (i) very weak sentiment (i.e., j j t 0) and (ii) low importance or low degree of correlation (i.e., r t 0). We call this region the stopword region. Therefore, to detect stopwords in our approach, we first build a SentiCircle for each word in the tweet corpus, calculate its overall Contextual semantics and sentiment by means of its SentiMedian, and check whether the word’s SentiMedian lies within the stopword region or not.
We assume the same stopword region boundary for all SentiCircles emerging from
the same Twitter corpus, or context. To compute these boundaries we first build the
SentiCircle of the complete corpus by merging all SentiCircles of each individual term
and then we plot the density distribution of the terms within the constructed SentiCircle.
The boundaries of the stopword region are delimited by an increase/decrease in the
density of terms along the X- and Y-axis. Table 1 shows the X and Y boundaries of the
stopword region for all Twitter datasets that we use in this work.
To evaluate our approach, we perform binary sentiment classification (positive / negative
classification of tweets) using a MaxEnt classifier and observe fluctuations (increases
and decreases) after removing stopwords on: the classification performance, measured in
terms of accuracy and F-measure, the size of the classifier’s feature space and the level
of data sparsity. To this end, we use 6 Twitter datasets: OMD, HCR, STS-Gold, SemEval,
WAP and GASP [
Figure 2 depicts the classification performance in accuracy and F1-measure as well as the reduction in the classifier’s features space obtained by applying our SentiCircle stopword removal methods on all datasets. As noted, our method outperforms the classic stopword list by 0.42% and 0.94% in accuracy and F1-measure on average respectively. Moreover, we observe that our method shrinks the feature space substantially by 48.34%, while the classic method has a reduction rate of 5.5% only.
Figure 3 shows the average impact of the SentiCircle and the classic methods on the sparsity degree of our datasets. We notice that our SentiCircle method always lowers the sparsity degree of all datasets by 1.17% on average compared to the classic method. 85 re 84 sau83 -eM‐82 F& 81 cya80 rcu79 cA78 77
Accuracy 50.99 40.99 e t a 30.99 R n o & 20.99 cu d e 10.99 R
In this paper we proposed a novel approach for generating context-aware stopword lists for sentiment analysis on Twitter. Our approach exploits the contextual semantics of words in order to capture their context and calculates their discrimination power accordingly. We have evaluated our approach for binary sentiment classification using 6 Twitter datasets. Results show that our stopword removal approach outperforms the classic method in terms of the sentiment classification performance and the reduction in both the classifier’s feature space and the dataset sparsity.
This work was supported by the EU-FP7 project SENSE4US (grant no. 611242).